Skip to content

CURRENT_STATE_LEDGER.md

Latest commit

 

History

History
155 lines (134 loc) · 7.73 KB

CURRENT_STATE_LEDGER.md

File metadata and controls

155 lines (134 loc) · 7.73 KB
Raw

Current State Ledger — SUPERSEDED

This ledger is superseded by /CURRENT_STATE_LEDGER.md (repo root), which tracks the effective state of the working tree after the rounds-8/9 Hermes live-streaming parity wave (suite: 1155 passed / 13 skipped / 0 failed; round-9 live grade: 39/39 phases, 0 failures).

The content below is the historical snapshot from the August 2026 engineering pass (794-test era) and is retained for claim-level context only. Per-phase closeout evidence remains in docs/PHASE_*_CLOSEOUT.md.

Historical snapshot (August 2026 engineering pass)

Living summary of what is verified in this repository and what remains open. Claim-level evidence lives in docs/PHASE_*_CLOSEOUT.md; this ledger tracks the effective state of the working tree after the August 2026 engineering pass.

Historical snapshot (August 2026 engineering pass)

Living summary of what is verified in this repository and what remains open. Claim-level evidence lives in docs/PHASE_*_CLOSEOUT.md; this ledger tracks the effective state of the working tree after the August 2026 engineering pass.

Round-8 security & correctness wave (2026-08-30)

A full-tree audit pass. Every finding was reproduced before it was fixed; pins live in tests/test_hardening_wave8.py (16 tests).

Credentials that were in the repository

  • A live Telegram bot token and a live provider API key were committed: the token appeared 1,342 times in realrun-evidence/server.log and 13 times in realrun-evidence/round5/engine.log (Bot API request URLs embed the credential), and both were hard-coded as module constants in scripts/e2e_round5_setup.py, realrun-evidence/env_common.py, realrun-evidence/s7_console_session.py, and realrun-evidence/s2_start_server.sh. A webhook secret was hard-coded in scripts/e2e_round5_engine.sh.
  • Fixed: every literal is now a required environment read that fails closed by name (E2E_BOT_TOKEN, E2E_PROVIDER_KEY, E2E_WEBHOOK_SECRET, REALRUN_BOT_TOKEN, REALRUN_API_KEY); the logs were redacted and untracked; .gitignore excludes *.log; and test_no_live_credentials_in_tracked_files scans every tracked text file for Telegram/OpenAI/AWS/GitHub/Slack credential shapes.
  • Operator action outstanding: the exposed values are in published git history. Revoke the bot token via @BotFather, rotate the provider key, and rewrite history if the old objects must be purged. Until then, treat both as compromised.

Other fixes

  • The admin GUI's CSRF token was sha256("csrf:" + session_id), so a leaked session id yielded the matching token; tokens are now random per session and dropped with the session.
  • DockerExecutor passed the full host environment to the docker CLI process; docker_cli_env() now passes only PATH, locale, and the daemon locators.
  • The auth middleware ran two synchronous SQLite reads on the event loop for every authenticated request; both now go through run_in_threadpool with the actor ContextVar re-bound in the worker thread.
  • Database never released cached connections, so ResourceWarning: unclosed database (CPython ≥ 3.13) failed unrelated tests non-deterministically under warnings-as-errors; a weakref.finalize now closes them. A test-side with sqlite3.connect(...) — which commits but does not close — was the second source.
  • Tests resolved the operator's real $ZERO_HOME, so the live owner_only access policy denied every interface intake (25 failures) and home-writing tests mutated the operator's config; a session-scoped fixture isolates it.
  • Suites driving a real loopback HTTP server now skip behind a loopback_http_works() probe instead of failing with ReadTimeout where loopback delivers requests but never returns responses.
  • Two order-dependent assertions were fixed: the polling-heartbeat tests waited a fixed 0.3s and asserted an iteration count (they now wait for the condition), and test_resolve_sink_path_from_env compared a POSIX-spelled path string on Windows (it now compares Path objects).
  • The package version contradicted the tree (pyproject.toml said 0.1.0 at v0.8.5, so the wheel, /healthz, --version, and the MCP handshake all reported a stale version). zero.__version__ is the single source of truth at 0.8.5; pyproject.toml reads it via [tool.setuptools.dynamic].
  • _now_utc_iso was defined 17 times with a byte-identical body; the canonical timestamp format now lives once in zero/app/clock.py.

Verified locally (deterministic, Windows host)

  • Full deterministic suite green after the round-8 wave: 1107 passed, 0 failed, 48 skipped (1155 collected) under ZERO_ENV=test on CPython 3.14. Skips are platform-gated POSIX cases, optional extras (tokenizer, psycopg), credential-gated live tests, and the loopback-HTTP suites that skip where the environment cannot complete a loopback round trip. The earlier figure was 794 passed / 31 skipped against a tree with 25 environment-induced failures.

Verified locally (deterministic, Windows host)

  • Full deterministic suite green: 794 passed, 0 failed, 31 skipped (825 collected) under ZERO_ENV=test. Skips are platform-gated POSIX cases, optional extras (tokenizer, psycopg), and credential-gated live tests that skip by design.
  • Ruff lint and format clean across src/, tests/, and scripts/.
  • HTTP API decomposed into zero.app.routers: sixteen per-domain router modules plus shared deps.py/models.py; app/api.py keeps application assembly only (304 lines). Golden route/OpenAPI tables prove the public surface identical across the move.
  • Release artifact gate green: wheel and sdist build from a clean committed tree; scripts/validate_release_artifacts.py exits 0 (all 32 migrations matched, required modules present, no unsafe paths, no credential matches).
  • Installed-console fail-closed contract holds: zero-develop serve without ZERO_ENV prints the exact configuration error and exits non-zero; the CI release job asserts this.
  • Fail-closed configuration boundary (ADR 0004) holds: structural validation (pool bounds) runs before capability gates; production mode still refuses missing secrets and development database paths.

Behavior fixes landed in the August 2026 engineering pass

  • Wizard group discovery crashed with NameError (probes.telegram_recent_chats referenced an undefined global); fixed with a real-HTTP regression test.
  • Pool-bound misconfiguration surfaced as an install hint when the [pg] extra was absent; structural validation now precedes the capability gate.
  • Release gate rejected every build once migrations 0029/0030 shipped (stale hand-maintained migration list); the expected set is now derived from the source tree, with a regression test covering both drift and silent empty-set failure modes.
  • CI release smoke invoked bare zero-develop, which dies in argparse now that subcommands are required; smoke steps use zero-develop serve.
  • Provider retry test used a wall-clock budget that flaked under full-suite load; it now asserts the recorded backoff sleeps directly.
  • $ZERO_HOME had seven duplicated resolvers; consolidated into zero.manage.core.config.zero_home (behavior unchanged).

Still open / not verified here

  • Live Telegram, OpenAI, and Anthropic integrations remain unverified (credential-gated suites skip unless explicitly enabled — see docs/LIVE_TESTING.md).
  • PostgreSQL integration tests require a disposable server via ZERO_TEST_PG_URL and were skipped; dialect unit tests ran.
  • Production rollout blockers in README.md ("Not production ready yet") and docs/PHASE_9_CLOSEOUT.md remain authoritative: this checkpoint is development-only.